Computer Aided Detection on Cervical Cancer from Unstained Cytological Images

Abstract

Cervical cancer is the fourth most common cancer in women and is mainly driven by high-risk HPV, which induces abnormalities in cervical cells. Early detection is crucial, but conventional Pap smear screening involves lengthy staining and subjective interpretation, affecting accuracy. This thesis develops a non-invasive, label-free, quantitative microscopy approach using Differential Interference Contrast (DIC) and Autofluorescence (AF) imaging for early cervical cancer detection. DIC offers detailed morphological information, while AF captures metabolic variations, providing complementary structural and biochemical insights. newline newlineThe thesis begins with global cervical cancer statistics and the significance of label-free imaging. The mathematical framework in Chapter 2 covers preprocessing, segmentation, feature extraction, feature selection, and classification. Since segmentation is essential, Chapter 3 introduces an overlapping cell segmentation method and proposes a Voronoi-based mixed-breed active contour (VMAC) model for accurate cell and nucleus detection. newline newlineChapter 4 classifies cervical cells using geometric and intensity features, with PCA for feature selection and SVM giving the best results. Chapter 5 extends classification to AF images using spectral texture features (SPTFs), where Gaussian SVM performs best. Chapter 6 applies deep learning (VGG16, ResNet50) for automated feature extraction, with ResNet50 outperforming VGG16. Chapter 7 presents an ensemble multimodal system integrating DIC, AF, handcrafted, and deep features, where bagging shows the highest accuracy. The thesis concludes by highlighting the complementary roles of DIC and AF in detecting early structural and metabolic changes. newline